Biases in the Blind Spot: Detecting What LLMs Fail to Mention
Iván ArcuschinUniversity of Buenos Aires [email protected], David ChaninUniversity College London [email protected] Adrià Garriga-AlonsoIndependentOana-Maria CamburuImperial College London · 2026
Read Biases in the Blind Spot: Detecting What LLMs Fail to Mention by Iván ArcuschinUniversity of Buenos Aires [email protected], David ChaninUniversity College London [email protected] Adrià Garriga-AlonsoIndependentOana-Maria CamburuImperial College London (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-10117.txt.
Description
Language models often explain themselves with plausible reasoning traces that hide the biases actually driving their choices. A 2026 paper led by Iván Arcuschin, with three co-authors, introduces a fully automated, black-box system for finding such hidden biases. Given a task data set, the system uses model raters to propose candidate bias concepts, then tests each on growing samples with positive and negative versions.
A concept counts as an unverbalized bias when it shifts performance a lot yet never appears as justification in the model's reasoning traces. Across seven models and three decision settings, hiring, loan approval, and university admissions, the system found previously unknown biases and confirmed hand-found ones, including gender, race, religion, and ethnicity.
Catalog record
| CREATOR | Iván ArcuschinUniversity of Buenos Aires [email protected], David ChaninUniversity College London [email protected] Adrià Garriga-AlonsoIndependentOana-Maria CamburuImperial College London |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 148 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-10117